Learning-Based Surrogate-Assisted Modelling Strategy Selection for Expensive Constrained Multi-Objective Optimisation
提出一种双Q学习代理辅助进化算法，动态选择三种建模策略，解决昂贵约束多目标优化问题，在基准测试和机械设计问题中验证了有效性。
Surrogate assisted evolutionary algorithms (SAEAs) have emerged as a promising approach to solving expensive constrained multi-objective optimisation problems (ECMOPs) by significantly reducing computational costs. However, existing SAEAs typically rely on fixed surrogate modelling strategies or manually predefined switching mechanisms, which limits their adaptability and effectiveness across diverse optimisation scenarios. To address this issue, this paper proposes a double Q-learning-based surrogate-assisted evolutionary algorithm, denoted as QSAEA, to solve expensive constrained multi-objective optimisation problems (ECMOPs). The algorithm combines three surrogate modelling strategies and adopts the double Q-learning mechanism for adaptive selection. The first strategy models only approximate the objectives, helping the population traverses across infeasible regions and approach the unconstrained Pareto front (UPF) quickly. The second strategy considers both objectives and constraint violation (CV), guiding the population moves towards the feasible region while preserving solution quality. The third strategy incorporates objectives and each unsatisfied constraint to perform local search within the feasible region, improving convergence accuracy and solution diversity. The double Q-learning mechanism takes the Pearson correlation between the contributions of individuals with respect to objective improvement and constraint violation as the state input. This allows the algorithm to dynamically choose the most suitable modelling strategy at each stage and adapt to the changing search landscape. Experimental results on a series of benchmark test suites and five real-world mechanical design problems demonstrate the effectiveness of QSAEA in solving ECMOP.